Parth
Parekh

I make machine learning work outside the lab: on factory floors, ships, phones and air-gapped servers.

Senior AI Engineer at QED Analyticals

150x
faster coal-blend optimization: a 5-hour genetic search down to under 2 minutes
92%
hull-number recognition for naval ship identification, on a Jetson AGX Orin
5.2x
faster judge model, distilled to 150M parameters, matching the 0.8B original on reading tasks

Right now

Training a small language model from scratch to measure which of DeepSeek's ideas, like mHC, Engram and MoE, still help at 33M parameters.
The LLM project
Published OpenJEV-Flash, a 150M-parameter judge model that matches the 0.8B original on reading tasks at 5x the speed.
OpenJEV-Flash
At QED Analyticals, running on-prem GenAI on two clustered NVIDIA DGX Spark nodes.
Experience

Selected work

All 23 projects

2026, systems
Air-Gapped Document Assistant
An assistant for engineering documents and drawings that runs entirely on local models: OCR, retrieval over a vector store and a knowledge graph, and an agent that cites its sources.
Fully air-gapped in production
How it works
2026, vision & edge
Foundry Vision and Traceability
A platform for a ductile iron pipe foundry that watches treatment stations on camera, ingests sensor, lab and historian data, and infers which metal became which pipe.
How it works
2025, models
Coal-Blend Optimizer
A genetic algorithm that finds the cheapest coal blend whose predicted coke still meets its targets. Scoring the whole population in one batched pass cut a run from 5 hours to under 2 minutes.
150x faster blend optimization
How it works
2026, models
OpenJEV-Flash
Alex Wortega's openjev judge, rebuilt on the 150M-parameter Ettin encoder and distilled from its 4B model. One cross-encoder call reranks, grades, checks for hallucinations or picks an option.
5.2x faster than openjev-0.8B
How it works
2024–26, vision & edge
Real-Time Mosquito Tracking
An edge video pipeline that finds, tracks and counts mosquitoes in a live camera feed for biosurveillance. It runs at 60 FPS on standard CPU hardware, with no GPU.
60 FPS on a CPU
How it works
2025–26, models
A DeepSeek-Style LLM from Scratch
A decoder-only language model written from scratch in PyTorch, with DeepSeek's V2 to V4.1 components as config switches, each measured against the same baseline.
Trains at 74% of the GPU's fp32 peak
How it works
2026, models
Compose Keyboard Swipe Neural Encoder
A 634K-parameter swipe-typing model for my Android keyboard, trained on 896K real swipes and run in plain Kotlin with no inference library. This site's home page runs the same weights in your browser.
85.8% top-1, up from 73%
How it works
2025–26, models
Sigma-Go Chess Engine
A chess engine built from scratch: a residual network suggests moves and judges positions, and Monte Carlo tree search looks 800 simulations ahead. It learns from human games, not self-play.
4.3x faster batched tree search
How it works
2026, systems
RustSQLite
A database engine written from scratch in Rust with only the standard library. It writes real SQLite 3 files that sqlite3 can open, and commits through a write-ahead log.
0 dependencies
How it works

Try a model I trained

The swipe-typing network from my Android keyboard, running in your browser. Draw across the keys and it reads the word from the shape of your path and its timing.

Try it

Swipe across the keys to spell a word.

Or type a word to see its path, and what the model reads

Loading the model, 2.5 MB

This is the real model from my Android keyboard: the same 2.5 MB encoder and 150,000-word lexicon, ported from Kotlin to TypeScript and running in your browser. About the swipe encoder

Experience

Roles in detail

Jul 2026 – now
Senior AI Engineer
QED Analyticals, Delhi, India · Hybrid
AI Infrastructure, GenAI Platforms, Production Recovery
Nov 2024 – Jul 2026
AI Engineer
QED Analyticals, Delhi, India · Hybrid
Industrial AI, Optimization, Edge Computer Vision
Mar 2024 – Dec 2024
System Software Engineer
Krishworks Technology & Research Labs, Bengaluru, India
IoT Infrastructure, Edge AI, Cloud OTA
May 2023 – Aug 2023
Project Lead
County of Simcoe, Simcoe County, Canada
Deep Learning, Spatial Simulation, Leadership
Jun 2022 – Dec 2022
System Software Engineer — Trainee
Krishworks Technology & Research Labs, Bengaluru, India
Mobile Systems, Fullstack CRM, BLE
Oct 2022 – Nov 2022
Software Engineer (Freelance)
LogicTerrain, Remote
App Performance, Android 12, Stability

On GitHub

596 contributions in the last year

github.com/xxparthparekhxx

ParthParekh